Even ideal artificial variants resist fine-tuning: LAMP shows sequence-to-function models still miss distal effects
Synopsis
LAMP injects artificial, strong and consistent transcription-factor-motif-perturbing SNVs into personal genome datasets to test whether sequence-to-function models can learn variant effects during fine-tuning. Even with the training set scaled to 12,111 genes and 12,011 such idealized causal SNVs, fine-tuned Enformer still trailed an elastic net baseline given only the SNV genotype on held-out genes and people. The analysis points to two obstacles: pre-trained priors that downweight distal regulatory elements make enhancer-to-gene links hard to strengthen, and sequence context differences between instances of the same motif obscure an otherwise easy-to-learn perturbation pattern.
Interpretation
After fine-tuning, sequence-to-function models still fail to learn a class of causal variant effects deliberately designed to be easy to learn, so the difficulty is not the complexity of natural genetic variation. Earlier post-training on personal genomes with paired expression values gave no meaningful improvement on held-out genes and people, and the reason was unclear. With STAT1-motif artificial SNVs injected into personal genomes from 670 GTEx whole-blood individuals and training scaled to 12,111 genes and 12,011 training SNVs, fine-tuned Enformer remained below an elastic net baseline using only the SNV genotype on held-out people across 88 test genes.
The model detects the artificial SNV in its convolutional layers, but that genotype information is not fully carried to the gene transcription start site, especially for distal motifs. The model learns SNVs placed near the transcription start site better than those at the window edges, and more data only partly closes the gap. Regressions on fine-tuned Enformer embeddings show that embeddings overlapping the SNV predict individual expression much better than the full forward pass, while transformer embeddings at the transcription start site carry markedly less information; the same gap appears across models for 45 transcription factor motifs.
Sequence context differences between instances of the same motif obscure the variant-induced motif perturbation pattern and reduce learning efficiency. Fine-tuning separately on 45 motifs produced widely different held-out accuracy, with the motif itself the main variable. Adding fixed flanking sequence to every candidate STAT1 motif let a model trained on 179 SNVs match one trained on about 12,000, with median held-out cross-individual correlation of 0.648 versus 0.643; across the 45 motifs, palindromicity and length-controlled information content had significant positive effects and explained 61% of the variance.
Perspective
The diagnosis covers variants that act by perturbing transcription factor motifs, uses GTEx whole-blood expression with Enformer as the main model, and checks Borzoi in some experiments; the authors state they did not test AlphaGenome with LAMP. They propose data augmentation that slides perturbed motifs within the sequence to break distance priors, and suggest a final fine-tuning step on unaltered data.
The authors note that fine-tuning on personal genomes with paired expression is highly compute intensive; LAMP simulates single-motif perturbation and does not cover variants acting through other mechanisms or regulatory logic shared across motifs; and because the variant effects are simulated, learning from real genetic variation is not guaranteed to behave identically. They also state that Borzoi and AlphaGenome reportedly still fall short on the personal genome task and retain priors biased away from distal elements, so they expect the findings to hold for them, but only Borzoi was tested in a subset of experiments.
